You are on page 1of 13

Analytical Uncertainty Propagation via Metamodels in

Simulation-Based Design under Uncertainty


Wei Chen*
Integrated DEsign Automation Laboratory (IDEAL), Northwestern University, Evanston, IL 60208-3111, USA

Ruichen Jin and Agus Sudjianto


Ford Motor Company, Dearborn, MI 48124, USA
In spite of the benefits, one of the most challenging issues for implementing optimization under uncertainty,
such as the use of robust design approach, is associated with the intensive computational demand of
uncertainty propagation, especially when the simulation programs are computationally expensive. In this
paper, an efficient approach to uncertainty propagation via the use of metamodels is presented. Metamodels,
created through computer simulations to replace expensive simulation programs, are widely used in
simulation-based design. Different from existing techniques that apply sample-based methods to metamodels
for uncertainty propagation, our method utilizes analytical derivations to eliminate the random errors as well
as to reduce the computational expenses of sampling. In this paper, we provide analytical formulations for
mean and variance evaluations via a variety of metamodels commonly used in engineering design
applications. The benefits of our proposed techniques are demonstrated through the robust design for
improving vehicle handling. In addition to the improved accuracy and efficiency, our proposed analytical
approach can greatly improve the convergence behavior of optimization under uncertainty.

Nomenclature
xR :

Random variables

pR ( x R ) :

Joint probability density function (PDF) of random variables

pi ( xi ) :
y (x) , y2 (x) :

Individual (marginal) probability density function

Bi ( x) :
hil ( xl ) :

Multivariate tensor-product basis function

Response mean and Variance


Univariate basis function

Nb :
C1,il , C2,i i l :
12

Number of multivariate basis functions


Univariate integrals involved in analytical uncertainty propagation

1.

INTRODUCTION

Development of efficient methods for uncertainty propagation has gained much attention in recent years due to
the increasing awareness of the importance of nondeterministic optimization. By uncertainty propagation, we mean
that the impact of input uncertainty on the variation of a model output (response) is studied. Despite the
advancement of methods for uncertainty propagation, the rising fidelity of engineering analyses has significantly
increased the computational expenses of simulation-based design and creates the barrier for applying
nondeterministic optimization to real engineering design applications.
One difficulty with the conventional statistical approach to uncertainty propagation is that it relies heavily on
the use of data sampling to generate probabilistic distributions of a system output. Monte Carlo simulation, a
random simulation based approach is very expensive. Even reduced sampling techniques, like Latin Hypercube
Sampling (Box et al. 1978) and Taguchis orthogonal array (Phadke 1989), can still require a large amount of
*

Associate Professor, Department of Mechanical Engineering, 2145 Sheridan Rd., Tech B224, Evanston, IL 602080311, weichen@northwestern.edu, Associate Fellow of AIAA.

Engineer

Technical manager
1
American Institute of Aeronautics and Astronautics

samples for complex systems. In practice, the mean-value based approximation methods (Parkinson, et al. 1993)
that use Taylor series expansions to estimate the first two moments (mean and variance) of a system response have
been used widely because of their simplicity. However, for nonlinear response functions, Taylor expansion is
generally not sufficiently accurate (Wu, et. al, 1990), especially when the parameter uncertainty is large. In recent
years, reliability analysis based approaches to uncertainty analysis have gained a lot of promises (Du and Chen
2003, Tu et al. 1999), where the concept of the Most Probable Point is utilized to generate the cumulative
distribution function of a model output via a serial of limit state values. They generally provide better accuracy
compared to mean-value based approximations and better efficiency compared to sampling-based methods.
However, for problems that require expensive high-fidelity engineering analyses that may take hours or even days
for a single simulation, the use of reliability analysis based methods is still not affordable.
Because the direct use of data sampling methods is very expensive, an alternative approach that is being widely
considered is the use of metamodels to replace the expensive simulation models (Chen et al. 1997; Delaurentis and
Mavris, 2000), either over the space of all variables or that of the random variables only. Data sampling is then
performed using metamodels. Studies show that when samples are used to estimate the performance mean and
variance in metamodel-based design optimization, the performance functions tend to be noisy causing problems for
optimization convergence (Padmanabhan and Batill, 2000). Some of the metamodeling techniques such as Kriging
(Sachs, et al., 1989) require more computational resources than the others (for example, polynomial response
surfaces). Applying sample-based methods could still be inefficient for these relatively expensive matamodels in
optimization under uncertainty
Our objective in this work is to develop an efficient approach to uncertainty propagation via the use of
metamodels. Different from existing techniques that apply sample-based methods to metamodels, our method
utilizes analytical derivations to reduce the computational expenses and to improve the accuracy by eliminating the
random errors of statistical sampling. Our focus in this work is on deriving the first two moments (mean and
variance) of a system response in optimization under uncertainty. Analytical formulations are provided for a variety
of commonly used metamodels, such as the polynomial (Box et al. 1978), the Kriging (Sacks, et al., 1989; Currin, et.
al, 1991), the Radial Basis Functions (Hardy, 1971; Dyn, et al., 1986), and the MARS (Friedman, 1991). We show
that even though the function forms of these metamodels vary significantly, they all follow the form of multivariate
tensor-product basis functions for which the analytical results of univariate integrals can be combined to evaluate the
multivariate integrals in uncertainty propagation. Our paper is organized as follows. In Section 2, we lay out the
general formulations for uncertainty propagation and introduce the concept of tensor-product basis functions. In
Section 3, we provide detailed analytical derivations for various types of metamodels. In Section 4, we illustrate our
approach and its advantages using the robust design for improving vehicle handing as an example. Section 5 is the
closure of this paper.

2. MATHEMATICAL BACKGROUND
2.1 Formulations in Uncertainty Propagation
Various formulations are used for optimization under uncertainty. Robust Design (Chen et al., 1996) is a widely
used paradigm where the emphasis is on improving the quality of a product through minimizing the effect of
variation (noises) without eliminating the causes. A detailed comparison of various nonlinear programming based
robust design formulations is provided in Du and Chen, 2002. As observed, the formulation selected is often a
tradeoff between the accuracy required for uncertainty assessments, e.g., probability of constraint being feasibility,
and the computational needs. In many practical applications, the first two moments of a performance distribution
(mean and variance) are widely used to form a robust design objective and to assess the constraint feasibility under
uncertainty.
Given a response y = f (x) , where x are input variables that can be deterministic or stochastic (i.e., subject to
randomness), associated with either design variables or noise variables, the response mean and variance are defined
by:

y ( x ) = f ( x ) pR ( x R ) d x R ,
2
American Institute of Aeronautics and Astronautics

(2.1)

y ( x) = [ f ( x) y ( x)] pR ( x R ) d x R ,
2

(2.2)

where R is the set of random variables. Here, it is assumed that the random variables are statistically independent,
i.e., their joint probability density function (PDF) is a product of individual or marginal PDF of all variables, i.e.,
pR ( x ) =
pl ( xl ) . Both the response mean and variance are functions of input variables x. When there are

l R

multiple random variables, the integrations in Eqs.1-2 involve multivariate integrals which are very computationally
expensive. As introduced in Section 1, applying advanced sampling approaches directly to a model for uncertainty
propagation is not practically feasible for problems with time-consuming model evaluations; random errors cannot
be avoided even applied to easy-to-compute metamodels. We find that if a function can be expressed as a tensorproduct basis function (c.f., Hastie, et al., 2001), then the analytical results of univariate integrals can be combined
to evaluate the multivariate integrals in Eqs.2.1-2.2 for uncertainty propagation. As it will be disclosed later, most
of the commonly used metamodels all follow the form of multivariate tensor-product basis functions for which the
formulations of mean and variance can be analytically derived. The concept of tensor product basis function is first
introduced next.
2.2 Tensor Product Basis Functions
A multivariate tensor-product basis functions Bi (x) is defined as a product of M univariate basis functions
hil ( xl ) , i.e.,
M

Bi (x) = hil ( xl ), i = 1, 2,..., N b .

(2.3)

l =1

For instance B(x)=x1x2 can be rewritten as B(x)=h1(x1)h2(x2), where h1(x1)= x1 and h2(x2)= x2 . Note here hil ( xl )
could be equal to 1. Then a special category of functions, called tensor-product basis functions, can be defined as a
linear expansion of these multivariate basis functions, i.e.,
Nb

Nb

i =1

i =1

l =1

f (x) = a0 + ai Bi ( x) = a0 + [ai hil ( xl )] ,

(2.4)

where ai (i =0, 1,..., Nb) are constant coefficients. Many commonly used metamodels, i.e., polynomial regression
model, MARS, RBF (with Gaussian basis functions), Kriging, can be expressed as tensor-product basis functions,
with details shown in Sections 3.2-3.5.
.

3. Analytical Formulations of Uncertainty Propagation via Metamodels


In this section, we first derive the generalized formulations for uncertainty propagation based on the concept of
tensor product basis functions. We then illustrate the analytical derivations of function mean and variance for a set
of commonly used metamodeling techniques.
3.1 Generalized UP Formulations in the Form of Tensor Product Basis Function
By substituting into Eq. 2.1 the expression of the tensor-product function (Eq.2.4), the performance mean as a
function of input variable x can be written as:
y ( x) =

f (x)

lR

Nb

= a0 +

[ pl ( xl )dxl ] =

Nb

a0 +

[a h ( x )] [ p ( x )dx ]
i

i =1

il

l =1

a [ h ( x ) p ( x )dx ] h ( x )
i =1

il

lR

il

lR

lR

(3.1)

Nb

= a0 +

[a C h ( x )],
1, il

i =1

lR

il

lR

where, C1,il is the mean of the univariate basis function hil ( xl ) , i.e.,
C1,il = hil ( xl ) pl ( xl ) dxl .

3
American Institute of Aeronautics and Astronautics

(3.2)

Similarly, if we substitute into Eq.2.2 the expressions of y (x) (Eq.3.1) and the tensor-product function
(Eq.2.4), we then derive the response variance as follows:
Y (x) =
2

Nb

Nb

[ a h ( x )] [ a C h ( x )] [ p ( x ) dx ]
i

il

i =1

Nb

1, il

i =1

l =1

lR

il

l R

lR

Nb

a h ( x )] [ p ( x ) dx ] [ a C h ( x )]

= [

i =1

Nb

Nb

i1 =1 i 2 =1

il

l =1

ai ai
1

i =1

l R

(C

1, i1l

[C

C1, i l ){
2

l R

1, il

2 , i1i 2 l

lR

il

l R

[h

/( C1, i l C1, i l )] 1}
1

(3.3)

l R

i1l

}.

( xl ) hi l ( xl )]
2

l R

It can be observed that both the mean and variance evaluations (Eqs.3.1 and 3.3) of a tensor product basis function
depend on two common sets of quantities defined by univariate integrals, i.e., the mean of univariate basis function
C1 and the inner product of two univariate basis functions C2, defined as:
C2,i1i2l = hi1l ( xl ) hi2l ( xl ) pl ( xl ) dxl .

(3.4)

Once the basis functions and the distributions of the variables are fixed, the values of C1 and C2 can be derived
analytically and used to evaluate the response mean/variance at any given design settings (i.e., for given values of
design input x). In the case that the distributions of several variables are changed, only the C1 and C2 corresponding
to these variables need to be re-evaluated. In the following subsections we illustrate the derivations of C1 and C2 for
various types of metamodels.
3.2 Derivations via Polynomial Metamodels
All polynomial regression models can be transformed into the tensor product form. The most widely used
second-order regression model is shown as:
M

f ( x) = 0 + i xi + ii xi + ij xi x j .
2

i =1

i =1

(3.5)

i =1 j = i +1

For the above model, for any 0i, jM and j0, we define multivariate basis functions as
B(i , j )

x j
2
= x j
x x
i j

i=0
i= j
i< j

(3.6)

The univariate basis functions corresponding to variable xl are:


1

h(i , j )l = xl
x 2
l

none of (i, j ) = l
only one of (i, j ) = l
both of (i, j ) = l

(3.7)

The polynomial model can be re-written as:


f ( x) = 0 +

0i , j M , j 0

ij B(i , j ) = 0 +

0 i , j M , j 0

ij h(i , j )l ,

(3.8)

l =1

where 0 j = j .
Based on Eqns. 3.2 and 3.4, the values of C1 and C2 are derived for the polynomial function, respectively as:
C1,(i , j )l

= xl pl ( xl )dxl = l
2
2
2
xl pl ( xl )dxl = l + l

none of (i, j ) = l
only one of (i, j ) = l
both of (i, j ) = l

4
American Institute of Aeronautics and Astronautics

(3.9)

C2,(i1 , j1 ) (i2 , j2 )l

xl pl ( xl )dxl = l

= xl 2 pl ( xl )dxl = l 2 + l 2
3
2
3
xl pl ( xl )dxl = l ,3 + 3l l + l

xl 4 pl ( xl )dxl = l ,4 + 4l l ,3 + 6l 2 l 2 + l 4

none of (i1 , j1 , i2 , j2 ) = l
only one of (i1 , j1 , i2 , j2 ) = l

(3.10)

only two of (i1 , j1 , i2 , j2 ) = l


only three of (i1 , j1 , i2 , j2 ) = l
all of (i1 , j1 , i2 , j2 ) = l

Here, l and l are the mean and variance of input variable xl; l , n (n = 3, 4) is the nth centered moment of xl,
2

i.e., l , n = ( xl l ) p ( xl ) dxl . The values of


n

l , l 2 , l ,3 , and l ,4 depend on the type of input distributions.

For uniform distributions, we can derive l = l / 3 , l , 3 = 0 , l , 4 = l / 5 ; for normal distributions,


2

we

obtain l , 3 = 0 , l , 4 = 3 l .
Noting that many items of C1 and C2 are equal to 1, we can further expand the mean (Eq. 3.1) and the variance
(Eq. 3.2) as:
4

y (x) = 0 + i i + ii i +
2

i R

i R

+ ( + ) x + x x ,
ij

i R j R , j i

i R

ij

j R

ij

(3.11)

i R j R , j i

y (x) = ii (i ,4 2i ) + ij i j + (i + ii i + ij j + ij xj )ii i ,3
2

iR

iR jR , j i

(i + ii i +

iR

iR

jR

jR

(3.12)

ij j + ij xj ) i ,
2

jR

jR

where ji = ij if i > j .

3.3 Derivations via Kriging Model


The Kriging method is a widely used metamodeling technique due to its capability of capturing nonlinear
behaviors. A Kriging model (Sacks, et al., 198; Currin, et al.,1999) with a constant global trend (i.e., Ordinary
Kriging) can be written as:
N
(3.13)
f (x) = + r (x)T R 1 (y ) = + r(x)T = + r (x) ,

i =1

i i

where = R 1 ( y 0 ) is a N1 vector (N is the number of sample points) and i is its ith element. The Kriging
model follows the form of a multivariate correlation function which can be defined to be a product of univariate
M

correlation functions, i.e., (t, u) = l (tl , ul ) , where l is the correlation function for lth dimension. The correlation
l =1

between x and a sample point xi, can be expressed as: ri (x) = (x, xi ) = l ( xl , xil ) . Let hil ( xl ) = l ( xl , xil ) , the Kriging
l =1

model is then transformed into a tensor-product function:


N

i =1

l =1

f (x) = + i hil ( xl ) .

(3.14)

Many different one-dimensional correlation functions are used with Kriging metamodeling (see, e.g., Currin, et.
al., 1991; Sacks, et. al., 1989). In this study, we focus on the commonly used Gaussian correlation function, with the
expression of hil ( xl ) = exp l ( xl xil ) 2

.
For uniform distributions of input variables, we can prove that (details are omitted here but can be found in Jin
2004):

5
American Institute of Aeronautics and Astronautics

C1,il =
C2,i1i2l =

1
2l

2l

1
2 l

(ail 2l ) (bil 2l ) ,
l

(3.15)

,
exp l ( xi1l xi2l )2 / 2 l (ai1l + ai2l ) l (bi1l + bi2l )

(3.16)

where ail = l + l x0il and bil = l l x0il ; (.) stands for the CDF of standard normal distributions.
For normal distributions of input variables, it can be proved that:
l
1
,
C1,il =
exp[
( l xil )2 ]
2
2
2 l l + 1
2 l l + 1
1

C2,i1i2l =

4l 2l + 1

(3.17)

exp 2 l
[(l xi1l )2 + (l xi2l )2 + 2l 2l (xi1l xi2l )2 ]
+

4
1
l l

(3.18)

Substituting the expressions of C1 and C2 derived here (Eqns. 3.17 and 3.18) into the generalized formulations
in Eqs.3.1 and 3.3, we then obtain the first and second moments in uncertainty propagation via Kriging models. The
results evaluated by this analytical approach are generally highly accurate. However, if the correlation matrix R is
ill-conditioned, large numerical errors could occur when l, l, or l are very small. This is due to the loss of

[C

significant digits in the evaluation of

[C

2, i1i2l

2, i1i2l

/(C1,i1l C1,i2l )] 1 in Eqn. 3.3: when l, l, or are very small,


l

/(C1,i1l C1,i2l )] will be very close to 1 (e.g., if = 0 , then C2,i i l = C1,i l C1,i l = 1 ); the subtraction of two nearly
12
1
2
l

equal numbers would cause the canceling error. This canceling error will be largely magnified by the deteriorated
accuracy of i due to the ill-conditioned R. Numerical approaches to reduce the canceling errors can be found in
Jins (2004).
3.4 Derivations via Gaussian RBF Model
The Gaussian Radial Basis Function model (also called RBFN model) can be written as,
N

y(x) = + ii (x) ,

(3.19)

i =1

where,

1
( x ) = exp
2

(x t
=
exp

(x

t il )

l =1

l =1

il

If

we

define

basis

functions

as

( x t )2
hil ( xl ) = exp l 2il , the RBF model is transformed into a tensor-product function:
2 i

i =1

l =1

f (x) = + [i hil ( xl )] .

(3.20)

Noting that the form of a Gaussian RBF model is similar to that of a Kriging model, except that the width

is

associated with each basis function instead of with each variable, the formulations can be derived in a similar way as
shown in Section 3.3.
For uniform distributions of inputs, it can be proved that

C1, il = i
{ ( l + l til ) / i ( l l til ) / i } ,
l 2
C2,i1i2l =

ai1i2
exp (t t )2 /(2i12 + 2i2 2 ) (l +l bi1i2l )/ ai1i2 (l l bi1i2l )/ ai1i2
l 2 i1l i2l

(3.21)

),

(3.22)

where, ai i = i i / i 2 + i 2 and bi i l = ( i 2ti l + i 2 ti l ) /( i 2 + i 2 ) .


12

12

For normal distributions of inputs, it can be proved that


C1,il =

l / l 2 + 1
2

exp[( l til )2 /(2 i 2 + 2 l 2 )]

6
American Institute of Aeronautics and Astronautics

(3.23)

i i

C2,i1i2 j =

bi1i2l

1
,
exp
[ i2 2 ( j ti1 j )2 + i1 2 ( j ti2 j ) 2 + j 2 (ti1 j ti2 j ) 2 ]
2bi1i2l

(3.24)

where bi i l = l 2 ( i 2 + i 2 ) + i 2 i 2 .
12

3.5 Derivations via MARS Model

A MARS model is a weighted sum of tensor spline basis functions, i.e.,


N
f ( x) = a + a B ( x) ,

(3.25)

i =1

where the tensor spline basis functions are defined as tensor products of univariate truncated power functions:
q
(3.26)
B (x) = [ s ( x t ) ] .

il

lK i

il

If we define a set of univariate basis functions:


1
hil ( xl ) =
q
[ sil ( xl til )]+

l Ki ,

(3.27)

l Ki

then the MARS model is transformed into a tensor-product function:


N
M
f ( x) = a + [ a
h ( x )] .

(3.28)

i =1

l =1

il

The values of C1 and C2 can be evaluated respectively by:


1

C1,il = q ub1
q
sil lb1 ( xl til ) pl ( xl )dxl

l Ki

(3.29)

l Ki

1
l K i1 , l K i2

(3.30)
l K i1 , l K i2
C1,i1l

C2,i1i2 l =
C1,i2l
l K i1 , l K i2

ub2

q
q
( si1l si2l ) lb2 [( xl ti1l )( xl ti2l )] pl ( xl )dxl l K i1 , l K i2
where ub1 = max( sil , til ) , lb1 = min( sil , til ) , ub = min[max(si l , ti l ),max(si l , ti l )] and lb = max[min(si l , ti l ),min(si l , ti l )] .
2
1
1
2
2
2
1
1
2
2

If lb ub2 , then C2,i i l = 0 . While the formulations above are general enough to be applied to different q, in the
following we focus on a commonly used setting, i.e., q=1.
2

12

For uniform distributions of input variables, it can be proved that:


s x2
C1,il = il [ l til xl ] ub
(l K i ) ,
lb
2 l 2
si l si l xl 3 (ti l + ti l ) xl 2
C2,i i l =
[
+ ti l ti l xl ] ub
(l K i , l K i ) ,
lb
2 l 3
2
where
and lb1a = max[lb1 , l l ] ;
ub1a = min[ub1 , l + l ]
f ( x) ba = f (b) f ( a ) ;

(3.31)

1a

1a

(3.32)

2a

12

2a

ub2 a = min[ub2 , l + l ]

and

lb2 a = max[lb2 , l l ] .

For normal distributions, it can be proved that:


t
x l
( x )2
1
C1,il = sil { l il ( l
)
exp[ l 2 l ]} ub
lb
l
l
2 l
2

C2,i1i2 l = si1l si2l {[

(l ti1l )(l ti2l )

+l ](

xl l

xl + l ti1l ti1l
2

exp[

(xl l )2
2l2

(l K i ) ,

2
]} ub
(l Ki1 , l Ki2 )
lb2

(3.33)
.

(3.34)

In the coming section, we demonstrate the advantages of the proposed method by applying it to an engineering
design problem under the robust design concept.
7
American Institute of Aeronautics and Astronautics

4. Robust Design for Improving Vehicle Handling


We use the robust deign for improving vehicle handling, in particular preventing rollover of ground vehicles, as
an example to illustrate the use of analytical uncertainty propagation (UP) in design under uncertainty. A topic
related to the analytical UP is the analytical method for global sensitivity analysis (GSA) to study the impact of
variations in input variables on the variation of a model output. The common underlying principle for both
activities is to combine the analytical results of univariate integrals to evaluate the multivariate integrals that are
associated with multvariate tensor-product basis functions (Jin et al. 2004). The details of the analytical approach to
GSA are provided in Chen et al. 2004, but will not be repeated here. As applying GSA plays an important role in
variable screening and building an accurate metamodel for further analytical UP in robust design, some of the results
from GSA are highlighted here.
4.1 Problem Description and Variable Screening

Rollover of ground vehicles is one of the major causes of highway accidents in the United States
(Mohemedshah and Council, 1997). To prevent vehicle rollover, we developed a robust design procedure (Chen, et
al, 2001) to optimize vehicle and suspension parameters so that the design is not only optimal against the worst
maneuver condition but is also robust with respect to a range of maneuver inputs. In our earlier work, the secondorder polynomial function was created as the response surface model; factor importance was examined by checking
the coefficients of the second-order polynomial function and screened based on the linear effects. In this work, our
analytical techniques for GSA are applied to the same problem; the Kriging model is built and used for deriving
analytical formulation for UP. Through comparisons, we illustrate the advantages of our newly developed methods.
The detailed description of the simulation program for studying the rollover behavior and the robust design
formulation for preventing rollover can be found in (Chen, et al, 2001). In brief, the rollover simulation is the
integrated computer tool ArcSim (Sayers and Riley, 1996) developed at the University of Michigan for simulating
and analyzing the dynamic behavior of 6-axle tractor/semi-trailers. Without building the metamodel, the testing of
an optimization scenario without robustness consideration takes at least five hours to converge (Michelena and Kim,
1998) on the Sun Ultra-1 workstation. When robustness considerations are introduced, the computational demand
becomes more significant.
Variables
HH1
KHX1
KT11
KT123
KT2123
LTS11
LTS123
LTS2123
M11
M123
M2123
SCFS11
SCFS123
SCFS2123

Table 1. Design Variables and Their Ranges


Description
Lower bound Upper bound
Height of hitch above ground
51.20
76.80
Hitch roll torsional stiffness
80000
120000
Axle 1 tire stiffness
5520.00
8280.00
Axles 2 & 3 tire stiffness
5520.00
8280.00
Axles 4,5,& 6 tire stiffness
4139.20
6208.80
Distance between springs on Axle 1
30.40
45.60
Distance between springs on Axles 2 & 3
30.40
45.60
Distance between springs on Axles 4,5 & 6
30.40
45.60
Laden load for Axle 1
11540
17310
Laden load for Axles 2 & 3
20358.40
30537.60
Laden load for Axle 4,5, & 6
16274.40
24411.60
Axle 1 spring stiffness scale factor
0.8
1.2
Axles 2 & 3 spring stiffness scale factor
0.8
1.2
Axle 4,5 & 6 spring stiffness scale factor
0.8
1.2

Unit
in
in-lb/deg
lb/in
lb/in
lb/in
in
in
in
lbm
lbm
lbm
/
/
/

Fourteen ArcSim input parameters corresponding to suspension and vehicle parameters are chosen as design
variables (control factors) (see Table 1). The steering and braking parameters are taken as the noise factors to
capture the range of maneuvering conditions (see Table 2). The level of braking (brake_level) is the amount of
braking pressure applied. The level of steering (steer_level) is the angle the steering wheel is turned. The starting
and ending times of braking defines when the driver starts and stops braking. The ending times of steering defines
when the driver stops steering. A rollover metric (R) is used as the response for which the metamodel is created.
8
American Institute of Aeronautics and Astronautics

The rollover metric is defined as the square root of the integral of the square of the rollover angle in a 5-second
period. If the rollover angle becomes greater than 45, rollover will inevitably occur.
Table 2. Noise Variables and Their Ranges
Variable
brake_level
start_brake
end_brake
steer_level
end_steer

Distribution
Uniform
Uniform
Uniform
Uniform
Uniform

Lower Bound
70
1.02
1.53
60
2.16

Upper Bound
120
1.38
2.07
100
3.24

Unit
psi
sec
sec
deg
sec

Based on the observations obtained from global sensitivity analysis (Chen et al. 2004), it is decided to use the
11 relatively important variables in subsequent procedures and freeze the eight relatively unimportant variables at
their nominal values. It is also found that three noise variables, i.e., brake_level, end_brake, and steer_level, play
very significant roles in the variability of the response. With the knowledge of variable importance, an additional
400 sample runs are applied in a sequential manner to build the Kriging model for the reduced set of 11 variables.
The R square value of the metamodel is at around 0.8 after confirmation. Previous experience has shown that this
problem is very nonlinear and it is difficult to obtain a metamodel with a very high accuracy (Chen et al., 2001; Jin
et al. 2000).
4.2 Robust Design Optimization

The design for preventing vehicle rollover under a range of maneuver conditions is modeled using the robust
design formulation as shown in Fig. 8. Note that the design variables here are the reduced set of design variables
after screening.
Given

Distribution of noise factors: start_brake, brake_level, end_brake, steer_level, and


end_steer (Table 2)
Kriging Model for the rollover metric

Find

Vehicle Parameters: HH1, LTS123, LTS2123, M11, M123, M2123


Satisfy
M11 + 2M123 + 3M2123 = Constant
(The total sum of the laden loads needs to be maintained as a constant.)
Boundaries of the variables
Maximize
Robust Design Metric
Cdk = (URL-y)/3y
y rollover metric; URL the critical roll over condition.
Figure 1. Robust Design Formulation

The design capability index Cdk, which measures the portion of the range of designs that satisfies the ranged
design requirement, is used as the robust design metric for improving vehicle handling performance. URL is the
upper requirement limited, i.e., the critical rollover condition, which is set as 45 deg-sec1/2 in this study. Statistically,
the use of 3 implies that when Cdk reaches 1, 99.865% of the performance distribution conforms to the
requirements, assuming that the performance is normally distributed.

9
American Institute of Aeronautics and Astronautics

Using the developed analytical techniques for uncertainty propagation (for assessing mean y and STD y ), in
particular based on the Kriging model in this case, we obtain the solutions to the robust design formulation in Fig. 1.
The baseline design and the robust design, together with the confirmed values of y, y, and Cdk (evaluated from the
results of 40 optimal LHS runs constructed for noise variables while fixing the design variables at their solutions)
are shown in Table 3. Graphical representations of the probability (based on the simulation results) of the rollover of
the baseline design and the robust design from our method are presented in Figs. 2 and 3. It is noted that while for
the baseline design, the probability of rollover is more than 12% (see Fig. 2); for the robust design obtained by using
our methods, no rollover occurs (see Fig. 3). The standard deviation of rollover metric for baseline, i.e., 18.29, is
much higher than those obtained from robust design formulation (3.15 for using our methods). In particular, we
observe that the distribution of rollover metric for the robust design is much farther away from the critical rollever
condition. Thus, we conclude that the robust design optimization has resulted in improved vehicle handling
performance.
Table 3. Solution of Robust Design
HH1*
LTS123*
LTS2123*
M11*
M123
M2123*
KHX1*
KT11
KT123
KT2123*
LTS11*
SCFS11*
SCFS123
SCFS2123

y
y

Cdk

Baseline Design
64.0
38.0
38.0
14425.0
25448.0
20343.0
1000000
6900.0
6900.0
5174.0
38.0
1
1
1
14.63
18.29
0.554

Robust Design
51.2
45.6
45.6
17310.0
30108.4
16274.4
1000000
6900.0
6900.0
5174.0
38.0
1
1
1
5.74
3.15
4.159

Robust Design in Chen et al (1999)


51.2
45.6
45.6
17310.0
25448.0
24411.6
1200000
5520.0
5520.0
4139.2
30.4
1.2
0.8
0.8
9.27
15.22
0.783

Note: The first six variables are the reduced set of design variables in our study. The variables with '*' are the reduced set of
design variable in the study of Chen et al (1999).

Mean: 14.63, STD: 18.29, Cdk: 0.554

0.45
0.4

Rollover

0.35
0.3
0.25
0.2
0.15
0.1
0.05
0

10

20

30

40
50
Rollover Metric

60

70

Figure 2. Rollover Metric Distribution for Baseline Design

10
American Institute of Aeronautics and Astronautics

Mean: 5.74, STD: 3.15, Cdk: 4.149

0.35
0.3

Rollover

0.25
0.2
0.15
0.1
0.05
0

10

20

30

40
50
Rollover Metric

60

70

Figure 3. Rollover Metric Distribution for Robust Design

Table 3 also provides the robust design obtained in our earlier study (Chen et al, 1999). It is noted that the
robust design achieved in this study is far better than the robust design result reported in the earlier work. Indeed, the
robust design procedure used in the earlier study fails to find a robust design with a Cdk large than 1 and rollover
occurs with probability close to 5% (Chen et al, 1999). For this reason, in the earlier study, the range of maneuver
condition was reduced to obtain a more desirable robust design. Specifically, the range of the three most significant
noise variables, i.e., brake_level, end_brake and steer_level, was reduced by 30% to achieve a robust design with Cdk
equal to 2.559. Nevertheless, even with this noise reduction, the robust design obtained is still not as good as the
robust design obtained in this study. Furthermore, the significantly better robust design solution is obtained in this
study with less number of simulations (720 in total including the simulation employed in screening, creating
metamodel, and confirmation versus 756 in earlier study).
Several factors have contributed to the improved robust design solution obtained by the proposed techniques in
this work. One major reason is that the Kriging model created based on space-filling sample points in this study is
much more accurate than the quadratic response surface model created based on ad hoc sampling approach in the
earlier work: The PRESS-R2 of quadratic metamodel, which estimates the prediction capacity and has similar
meaning as the R2CV, is merely 0.504, while the R2CV of the metamodel in this study is 0.792. Furthermore, the global
sensitivity analysis method used in this work has identified M123 as a critical variable through TSI due to its strong
interactions with noise variables. Chen's et al. earlier study, however, fails to identify M123 as an important
variable because a quadratic response model was created for the screening purpose, only the linear part of main
effect was used in ranking variable importance. One more reason behind the improved robust design solution is that
in this study, analytical uncertain propagation is used, which provides more accurate evaluation of performance
mean and variance than the 3-level full factorial design (243 sample points) approach used in Chen et al's work.
To further illustrate the benefits of using analytical UP compared to applying Monte Carlo samples to the
metamodel, we illustrate in Figs. 4-7 a comparison between the response mean and response variation obtained from
these two methods, respectively. The pictures show the results from the analytical method and the Monte Carlo
method (1000 random sample points), when all the design variables except M123 and LTS 123 are fixed at their
nominal values. From the figures, it is observed that the Monte Carlo results for both mean and STD are very noisy,
which will cause serious convergence problem in the process of optimization. We tested with different starting
points for using Monte Carlo samples for UP in the optimization process. In all cases, the optimizer fails to
converge. Furthermore, the Monte Carlo method is far more expensive than analytical method. For instance, the
evaluation of 2121grid points in the graphs takes 139 seconds for Monte Carlo method, while it takes only 38
seconds (PIII 650MHZ, MATLAB) when using the proposed analytical method.

11
American Institute of Aeronautics and Astronautics

Analytical Method

Monte Carlo Method

25

Rollover Metric Mean

Rollover Metric Mean

25

20

15

10

5
2

15

10

5
2

30
2.5

20

35
3

35

x 10

45

45
LTS123

M123

Figure 4. Mean of Rollover Metric


by Analytical Method

Figure 5. Mean of Rollover Metric


by Monte Carlo Method

Analytical Method

Monte Carlo Method

25

25

20

20

Rollover Metric STD

Rollover Metric STD

40

LTS123

M123

30

2.5

40

x 10

15

10

5
2

15

10

5
2
2.5
4

x 10

3
M123

45

40

35

2.5

30

x 10

LTS123

45

M123

Figure 6. STD of Rollover Metric by Analytical


Method

40

35

30

LTS123

Figure 7. STD of Rollover Metric by Monte Carlo

5. Closure
The fundamental contribution of this work is the development of analytical techniques for assessing the mean
and variance of a model output via the use of metamodels in simulation-based design under uncertainty. We
illustrate that the commonly used metamodels such as polynomial, Kriging, the Radial Basis Functions, and MARS
all follow the form of multivariate tensor-product basis functions for which the analytical results of univariate
integrals can be combined to evaluate the multivariate integrals in uncertainty propagation.
Through the example problem, we demonstrate that compared to the existing sampling based approaches to
uncertainty propagation, our approach provides more accurate as well as more efficient uncertainty propagation
results. The techniques are especially useful for design applications that require computationally expensive
simulations, where the metamdoels are often readily available to designers. Analytical uncertainty propagation
reduces the noises associated with sampling methods and greatly facilitates the convergence of robust design
optimization.
As a part of our research effort, the same idea presented in this paper has been used for deriving analytical
formulations for global sensitivity analysis (Chen et al. 2004). The knowledge obtained though global sensitivity
analysis (Saltelli et al. 1997) offers insights into the model behavior, provides guidance in reducing the problem
size, and helps to identify sources for variance reduction. Overall, they help designers to make informed decisions
in product design with the consideration of uncertainty. Future work in this area will involve the consideration of
the dependence of the variability of input variables and higher moments in uncertainty propagation.

Acknowledgments
The support from Ford URP (University Research Program) and NSF grant DMI0335877 is greatly
acknowledged. We are grateful for the Arcsim program provided by the Automotive Research Center (ARC) at
University of Michigan.
12
American Institute of Aeronautics and Astronautics

References
Box, G. E. P., Hunter, W. G., and Hunter, J. S., 1978, Statistics for Experimenters, Wiley, New York.
Chen, W., Garimella, R., and Michelena, N., 2001, Robust Design for Improved Vehicle Handling Under a Range
of Maneuver Conditions, Engineering Optimization, Vol. 33, No 3, 303-326.
Chen, W., Jin, R., and Sudjianto, A., 2004, Analytical Variance-Based Global Sensitivity Analysis in SimulationBased Design under Uncertainty, DETC2004-57484, 2004 ASME Design Automation Conference, Salt Lake
City, UT, September 28-October 2, 2004.
Chen, W., Allen, J. K., Schrage, D.P., and Mistree, F., 1997, "Statistical Experimentation Methods for Achieving
Affordable Concurrent Design", AIAA Journal, 35 (5), 893-900.
Chen, W., Allen, J.K., Mistree, F., and Tsui, K.-L., 1996, A Procedure for Robust Design: Minimizing Variations
Caused by Noise Factors and Control Factors, ASME Journal of Mechanical Design, 18(4), pp. 478-485.
Currin, C., Mitchell, T., Morris, M.D., and Ylvisaker, D., 1991, Bayesian prediction of deterministic functions, with
applications to the design and analysis of computer experiments, Journal of the American Statistical
Association, 86 (416).
DeLaurentis D. and Marvris D., 2000, Uncertainty Modeling and Management in Multidisciplinary Analysis and
Synthesis, The 38th AIAA Aerospace Sciences Meeting, Paper No. AIAA 2000-0422, 10-13 January 200, Reno,
NV.
Frangopol, D. M., 1985, Structural Optimization Using Reliability Concepts, Journal of Structural Engineering,
ASCE, 111, 11, 2288-2301.
Du, X, and Chen, W., 2004, "Sequential Optimization and Reliability Assessment Method for Efficient Probabilistic
Design" ASME Journal of Mechanical Design, 126(2), pp. 225-233, March 2004.
Du, X. and Chen, W., 2000, Towards a Better Understanding of Modeling Feasibility Robustness in Engineering,
ASME Journal of Mechanical Design, 122(4), pp. 357-583.
Dyn, N., Levin, D. and Rippa, S., 1986, Numerical procedures for surface fitting of scattered data by radial basis
functions, SIAM Journal of Scientific and Statistical Computing, 7(2), pp. 639-659.
Friedman, J. H., 1991, Multivariate Adaptive Regression Splines, The Annals of Statistics, 19(1), pp. 1-141.
Hardy, R.L., 1971, Multiquadratic equations of topography and other irregular surfaces, J. Geophys. Res., 76, pp.
1905-1915.
Hastie, T., Tibshirani, R., Friedman, J., 2001, The Elements of Statistical Learning: data mining, inference, and
prediction, Springer-Verlag, New York
Jin, R., 2004 Enhancements of Metamodeling Techniques in Engineering Design, PhD dissertation, University of
Illinois at Chicago, Chicago, IL., June 2004.
Jin, R, Chen, W., and Simpson T., 2001, Comparative Studies of Metamodeling Techniques under Multiple
Modeling Criteria, Journal of Structural & Multidisciplinary Optimization, 23(1), 1-13.
Michelena, N. and Kim, H-M, 1998, Worst-Case Design for Vehicle Handling, Fourth Annual Conference on
Modeling and Simulation of Ground Vehicles, The Automotive Research Center, The University of Michigan,
Ann Arbor, May 19-20, 1998.
Mohamedshah, Y. and Council, F., 1997, Synthesis of Rollover Research, March 10, 1997.
Padmanabhan, D., Batill, S.M., 2000, An Iterative Concurrent Subspace Robust Design Framework, Proceedings
of the 8th AIAA/USAF/NASA/ISSMO Symposium on Multidisciplinary Analysis & Optimization, Sep, Paper No.
AIAA 2000-4841, Long Beach, CA.
Parkinson, A., Sorensen, C., and Pourhassan, N., 1993, A General Approach for Robust Optimal Design, ASME
Journal of Mechanical Design, 115(1), pp.74-80.
Phadke, M.S., 1989, Quality Engineering Using Robust Design, Prentice Hall, Englewood Cliffs, NJ.
Sacks, J., Welch, W. J., Mitchell, T. J. and Wynn, H. P., 1989, Design and Analysis of Computer Experiments,
Statistical Science, Vol. 4, No. 4, pp. 409-435.
Saltelli, A., Tarantola, S. and Chan, K., 1999, A Quantitative Model-Independent Method for Global Sensitivity
Analysis of Model Output, Technometrics, 41(1), 39~56
Sayers, M.W. and Riley, S.M., 1996, Modeling Assumptions for Realistic Multibody Simulations of the Yaw and
Roll Behavior of Heavy Trucks, SAE Paper No. 960173. Society of Automotive Engineers, Warrendale, PA.
Tu, J., Choi, K.K and Young H.P., 1999, A New Study on Reliability-Based Design Optimization, ASME Journal
of Mechanical Engineering, 121(4), pp. 557-564.
Wu, Y.-T., Millwater, H.R., and Cruse T. A., 1990, An Advance Probabilistic Analysis Method for Implicit
Performance Function, AIAA Journal, Vol. 28, pp. 1663-1669.

13
American Institute of Aeronautics and Astronautics

You might also like